Accurate estimation of prediction models for operator-induced defects in assembly manufacturing processes

The presence of defects in industrial manufacturing may compromise the final quality and cost of a product. Among all possible defect causes, human errors have significant effects on the performances of assembly systems. Much research has been conducted in recent years focusing on the problem of defect generation in assembly processes, considering the close connection between assembly complexity and human errors. It was observed that the relationship between the average number of defects introduced during each assembly phase and the related assembly complexity follows a power-law relationship. Accordingly, many authors proposed a data logarithmic transformation in order to linearize the mathematical model. However, as has already been discussed in literature, when the model is retransformed in the original form a significant bias may occur, leading to completely wrong predictions. In this paper, the bias due to the logarithmic transformation of models for predicting defects in assembly is analyzed and discussed. Two alternative methods are proposed and compared to overcome this drawback: the use of a bias correction factor to the retransformed fitted values and a power-law nonlinear regression model. The latter has proved to be the best approach to predict defects with few non-repeated data and affected by high variability, such as in the case under study.